Would it be possible to see your per-GPU, per-game data that you've recorded to get these numbers? I'm sure a lot of people would be curious to know the details behind the percentages. Also, specifying which games are demos\trials can be pretty important since there can be a big difference in stability and features from early demos to fully patched up versions of games.

Thanks for the testing. Looking forward to reading more about it.

Any particular reason for using first unified Kepler driver for GTX 670, instead of last (306.81) ?
Note : 301 driver doesn't have GTX 660-650 cards listed as supported as well.

Judging by that list, 7970 has some issues related to 3 Gb of memory.

A few of the games put out an explicit VRAM error. For Moto Racer 3 GB seems to be a problem, 4 GB for Thief (Gold). I'd stay under 2 GB to be on the safe side.

The Serpent Rider wrote on 2025-11-04, 21:54:

Judging by that list, 7970 has some issues related to 3 Gb of memory.

2GB would be same with HD 78xx or HD77xx and DX3+, like HD 7750 1GB?

vvbee wrote on 2025-11-04, 03:53:

Games: Darkstone, Delta Force 2, Die Hard Trilogy, F1 2000, Formula 1, Gothic 2, Homeworld, Johnny Herbert's Grand Prix World Championship, MechWarrior 3, Monster Truck Madness, Monster Truck Madness 2, Motocross Madness, Moto Racer, Motorhead, Project IGI, Rainbow Six, Take No Prisoners, Thief, Thief 2, Virtua Fighter 2 (Direct3D), Warhammer: Dark Omen.

Of those games, Gothic 2 ran perfectly on my GTX 980 Ti. Thief 2 had color banding issues due to the lack of 16-bit dithering, but otherwise ran fine on that card as well. The other titles I'm not very familiar with, or haven't tried on that system.

vvbee wrote on 2025-11-04, 03:53:

Some of these are trial versions.

I wouldn't recommend using game demos for evaluating compatibility. Those are often based on old, pre-release code and don't have the fixes and optimizations from the retail version and official patches.

vvbee wrote on 2025-11-04, 03:53:

We can predict that if you have Kepler or Maxwell in your XP build you're more likely against the idea of playing older games in XP.

A fair assumption. WinXP means a lot of different things to different people. For me, it's reminiscent of the 2002-2007 era of gaming, give or take a year. So my focus is mostly on DX8 and DX9 titles, which usually run fine on Maxwell cards. For the sake of comparison, here's a list of games which generally worked well on my GTX 980 Ti under WinXP, though some of them needed a few tweaks.

Demos are even better since by definition the most compatible card needs the fewest changes.

Generally, playing DX5 or lower games directly on DX10+ is pointless anyway, because there's no 16-bit dithering and it looks pretty bad. As well as quite a large chunk of DX6 games.

vvbee wrote on 2025-11-05, 12:15:

Demos are even better since by definition the most compatible card needs the fewest changes.

I guess whether it is better depends on the goal of all the testing. If the goal is to help people determine which graphics card to use in their XP build to get the best experience in the games they want to play, then using the versions that people will actually play would be better. If the goal is to prove a point or create some kind of compatibility scoring system for GPUs, then yeah, using the most finicky and least refined or updated versions of games will exaggerate flaws and potentially give show clearer differences between GPUs.

I've brought this up in other threads in recent years, but the video chip DOS compatibility charts are a good example of how interesting data can, sometimes unintentionally, end up being generalized into a scoring system. I don't hear about it so much anymore, but 5-10 years ago people were kind of obsessing over which chips in that chart had the least "red" when talking about what to get for a build. You'd hear about how a certain card was worthless for DOS because of the chart, or how another one is the gold standard, basically because it was green the whole way across. In reality, many of us may not have played 80% of the games on that chart, or any of the most problematic ones. Like... how many VOGONS members are actually playing Mario Shareware in DOS? Funny enough, I had found other games that were less popular than something like Doom or Quake, but were actual full games that people might want to play, and they had issues with certain S3 VirgeDX gold-standard cards, while a Matrox Millennium (which looks positively dismal on the chart) worked flawlessly.

Anyway, just wanted to throw that out there. If you're looking for the cards with the worst compatibility, using the least-refined versions of games and demos will definitely give you results. If you're looking for the best compatibility for people who are building retro PCs, using the most refined versions of games that people will play would give more accurate results. If 9 out of 10 game demos from 1998 were broken on a given GPU, that would be an indicator of an issue. If the fully patched versions of those games all work fine however, then that changes how the data should be used.

If your experience with a card in general reflects my rating of it then I've sampled well and the list is validated. Whether a particular game would work is out of scope, hence why you don't see per-game results.

16-bit is a matter of taste. Dithering can look ugly in itself, and there are various implementations of it. This isn't an image quality comparison though except in the sense of render glitches.

vvbee wrote on 2025-11-05, 21:27:

If your experience with a card in general reflects my rating of it then I've sampled well and the list is validated. Whether a particular game would work is out of scope, hence why you don't see per-game results.

16-bit is a matter of taste. Dithering can look ugly in itself, and there are various implementations of it. This isn't an image quality comparison though except in the sense of render glitches.

If at all possible, having a list of the versions (full\demo and version number) of the games tested listed in the OP would be very helpful for those wanting to make use of this.

I think we can predict some trends now. Radeons will do DX5 as a general rule and some go down to DX3. Pre-Kepler GeForces have acceptable reliability down to DX6. Later GeForces are more of a footnote. At this point there's no GeForce 6000 series card tested, but the 7000 series Quadro is mediocre so I don't expect much. Different drivers could have some impact one way or the other. There are always individual games that will or won't work, but that's not the focus here.

The HD 7750 has some secret sauce, and the HD 3450 is an outlier in this direction as well. In fact the 7750 is compatible with DX3-11, can't say that about many cards. Whether other Radeon HD 7000 series cards have the same compatibility isn't clear. The 7970 says no, but 2+ GB of VRAM reduces compatibility to some extent.

The Matrox P690 is an outlier in being equally hit or miss in every DX version.

vvbee wrote on 2025-11-07, 19:31:

The HD 7750 has some secret sauce, and the HD 3450 is an outlier in this direction as well. In fact the 7750 is compatible with DX3-11, can't say that about many cards. Whether other Radeon HD 7000 series cards have the same compatibility isn't clear. The 7970 says no, but 2+ GB of VRAM reduces compatibility to some extent.

If this a thing with 2Gb or more specifically, there's a bunch of Radeon 7850 models with 1 Gb. Basically 2x performance of 7750 with the same driver model.

The Serpent Rider wrote on Yesterday, 16:25:

vvbee wrote on 2025-11-07, 19:31:

The HD 7750 has some secret sauce, and the HD 3450 is an outlier in this direction as well. In fact the 7750 is compatible with DX3-11, can't say that about many cards. Whether other Radeon HD 7000 series cards have the same compatibility isn't clear. The 7970 says no, but 2+ GB of VRAM reduces compatibility to some extent.

If this a thing with 2Gb or more specifically, there's a bunch of Radeon 7850 models with 1 Gb. Basically 2x performance of 7750 with the same driver model.

Funny you mention that, I just grabbed a complete in-box HD 7850 1GB for almost nothing. I will be curious to see how it performs in an XP system and how far back the games work without issues. Also very curious to see how it compares in performance and compatibility for old games to something like a GTX 750 Ti or GTX 660.

I don't see any GTX 660 models with 1 gigabyte, so it should be absolutely identical to GTX 670. As for 750 Ti, it probably will be better with 1 gigabyte, but judging by another Maxwell 1.0 card in the list (GTX 745), the compatibility is already too broken.

The Serpent Rider wrote on Today, 00:09:

I don't see any GTX 660 models with 1 gigabyte, so it should be absolutely identical to GTX 670. As for 750 Ti, it probably will be better with 1 gigabyte, but judging by another Maxwell 1.0 card in the list (GTX 745), the compatibility is already too broken.

Right, I was just saying that it'll be interesting to compare them to popular Nvidia cards of the era to see how much of a difference I find (in whatever games I test) considering the memory and architecture differences.

vvbee wrote on 2025-11-04, 03:53:

This is where I test various PCIe graphics cards' compatibility with DX3-7 under XP. […]
Show full quote

This is where I test various PCIe graphics cards' compatibility with DX3-7 under XP.

Games: Darkstone, Delta Force 2, Die Hard Trilogy, F1 2000, Formula 1, Gothic 2, Homeworld, Johnny Herbert's Grand Prix World Championship, MechWarrior 3, Monster Truck Madness, Monster Truck Madness 2, Motocross Madness, Moto Racer, Motorhead, Project IGI, Rainbow Six, Take No Prisoners, Thief, Thief 2, Virtua Fighter 2 (Direct3D), Warhammer: Dark Omen. Some of these are trial versions.

Results so far:

Copy code to clipboard

1Card                        Score  Rating  Driver
2-------------------------------------------------
3Matrox G550 PCIe (32 MB)    83%    DX3+    5.91
4Radeon X300 (256 MB)        83%    DX3+    8.12
5Radeon HD 7750 (1 GB)       71%    DX3+    iCafe
6Radeon HD 3450 (512 MB)     69%    DX3+    8.12
7Radeon HD 2600 PRO (512 MB) 66%    DX3+    8.12
8Radeon HD 3870 (512 MB)     60%    DX3+    8.12
9Nvidia NVS 300 (512 MB)     59%    DX6+    260.99
10Quadro FX 3800 (1 GB)       59%    DX6+    260.99
11Radeon HD 5770 (1 GB)       59%    DX5+    iCafe
12Radeon R7 360 (2 GB)        57%    DX5+    iCafe
13Radeon HD 7970 GHz (3 GB)   54%    DX5+    iCafe
14GeForce GTX 560 (1 GB)      54%    DX6+    275.33
15Quadro FX 1500 (256 MB)     54%    DX6+    94.24
16GeForce GTX 460 (768 MB)    52%    DX6+    260.99
17Matrox P690 (128 MB)        50%    n/a     2.06
18GeForce GT 640 (GF116 3 GB) 35%    DX6+    275.33
19GeForce GTX 670 (2 GB)      28%    DX7+    301.42
20GeForce GTX 980 (4 GB)      19%    DX7+    344.11
21GeForce GTX 745 (4 GB)      16%    DX7+    340.52

The G550 is a baseline for how compatible the games are with hardware rendering under XP. But its render speed is generally too low.

We can predict that if you have Kepler or Maxwell in your XP build you're more likely against the idea of playing older games in XP.

It'd be interesting to see how the X300 scores when more compatible drivers like 6.2 and 7.11 are used.